from typing import List, Dict, Any
import asyncio
from langchain_openai import ChatOpenAI
from langchain.prompts import PromptTemplate
from langchain.callbacks import AsyncIteratorCallbackHandler
from services.vector_store import AsyncVectorStore
from utils.logger import get_logger
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import os
import time
import torch
import json
class AsyncRAGService:
"""异步 RAG 服务主类"""
def __init__(self):
self.logger = get_logger(__name__)
self.vector_store = AsyncVectorStore()
self.openai_api_base = os.getenv("OPENAI_BASE_URL")
self.openai_api_key = os.getenv("OPENAI_API_KEY")
self.rerank_model_path = os.getenv("RERANK_MODEL_PATH", "")
self.open_rerank = bool(self.rerank_model_path)
self.llm = ChatOpenAI(
model="deepseek-r1:8b",
temperature=0.7,
openai_api_key=self.openai_api_key,
openai_api_base=self.openai_api_base,
)
self.prompt_template = PromptTemplate(
input_variables=["context", "question"],
template=(
"基于以下上下文回答问题。如果上下文中没有相关信息,请说明无法从提供的文档中找到答案。\n\n"
"上下文:\n"
"{context}\n\n"
"问题:{question}\n\n"
"答案:"
),
)
# 如果需要重排
if self.open_rerank:
self.logger.info("初始化 Ranker 模型...")
self.tokenizer = AutoTokenizer.from_pretrained(
self.rerank_model_path, padding_side="left"
)
# 强制设置 padding token
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
self.rerank_model = AutoModelForSequenceClassification.from_pretrained(
self.rerank_model_path
).eval()
# 确保模型配置与 tokenizer 一致
if hasattr(self.rerank_model.config, "pad_token_id"):
self.rerank_model.config.pad_token_id = self.tokenizer.pad_token_id
self.logger.info("✓ 初始化 Ranker 模型成功")
self.logger.info("RAG服务初始化完成")
async def add_document_async(self, content: str, filename: str) -> str:
"""异步添加文档"""
start_time = time.time()
try:
self.logger.info(f"开始添加文档: {filename}")
result = await self.vector_store.add_document_async(content, filename)
duration = time.time() - start_time
self.logger.info(f"文档添加成功: {filename}, 耗时: {duration:.2f}s")
return result
except Exception as e:
duration = time.time() - start_time
self.logger.error(
f"文档添加失败: {filename}, 错误: {str(e)}, 耗时: {duration:.2f}s"
)
raise
async def chat_async(
self, question: str, top_k: int = 3, temperature: float = 0.7
) -> Dict[str, Any]:
"""异步聊天问答"""
start_time = time.time()
try:
self.logger.info(f"开始处理问答: {question[:50]}...")
# 异步检索相关文档
search_results = await self.vector_store.search_async(question, top_k)
self.logger.debug(f"检索到 {len(search_results)} 个相关文档")
if not search_results:
self.logger.warning("未找到相关文档")
return {
"answer": "抱歉,我无法在现有文档中找到相关信息来回答您的问题。",
"sources": [],
"processing_time": time.time() - start_time,
}
# rerank
reranked_results = await self._rerank_results(question, search_results, skip_rerank=not self.open_rerank)
# 并行执行上下文构建和 LLM 调用准备
context_task = asyncio.create_task(
self._build_context_async(reranked_results)
)
sources_task = asyncio.create_task(
self._format_sources_async(reranked_results)
)
# 等待上下文构建完成
context = await context_task
# 异步生成回答
self.llm.temperature = temperature
prompt = self.prompt_template.format(context=context, question=question)
response = await asyncio.to_thread(self.llm.invoke, prompt)
# 等待源信息格式化完成
sources = await sources_task
duration = time.time() - start_time
self.logger.info(f"问答处理完成, 耗时: {duration:.2f}s")
return {
"answer": response.content,
"sources": sources,
"processing_time": duration,
}
except Exception as e:
duration = time.time() - start_time
self.logger.error(f"问答处理失败: {str(e)}, 耗时: {duration:.2f}s")
raise
async def chat_stream_async(
self, question: str, top_k: int = 3, temperature: float = 0.7
):
"""异步流式聊天问答"""
start_time = time.time()
# 异步检索相关文档
search_results = await self.vector_store.search_async(question, top_k)
if not search_results:
yield {
"content": "抱歉,我无法在现有文档中找到相关信息来回答您的问题。",
"is_final": True,
"sources": [],
"processing_time": time.time() - start_time,
}
return
# rerank
reranked_results = await self._rerank_results(question, search_results, skip_rerank=not self.open_rerank)
# 构建上下文和源信息
context_task = self._build_context_async(reranked_results)
sources_task = self._format_sources_async(reranked_results)
context = await context_task
# 设置 LLM 参数
prompt = self.prompt_template.format(context=context, question=question)
callback_handler = AsyncIteratorCallbackHandler()
stream_llm = ChatOpenAI(
model="deepseek-r1:8b",
streaming=True,
callbacks=[callback_handler],
openai_api_key=self.openai_api_key,
openai_api_base=self.openai_api_base,
)
self.logger.info("启动 LLM 流式生成任务...")
task = asyncio.create_task(stream_llm.ainvoke(prompt))
self.logger.info("LLM 流式生成任务已启动")
async for token in callback_handler.aiter():
yield {
"content": token,
"is_final": False,
"sources": None,
"processing_time": None,
}
await task
# 最后一个数据块包含完整信息
sources = await sources_task
yield {
"content": "",
"is_final": True,
"sources": sources,
"processing_time": time.time() - start_time,
}
async def get_documents_async(self) -> List[Dict[str, Any]]:
"""异步获取文档列表"""
return await self.vector_store.get_documents_async()
async def delete_document_async(self, doc_id: str) -> bool:
"""异步删除文档"""
return await self.vector_store.delete_document_async(doc_id)
async def _format_sources_async(
self, search_results: List[Dict[str, Any]]
) -> List[Dict[str, Any]]:
def _format_sources():
return [
{
"filename": r["metadata"]["filename"],
"content": (
(r["content"][:200] + "...")
if len(r["content"]) > 200
else r["content"]
),
"similarity": 1 - r["distance"],
"rerank_score": r.get("rerank_score", None),
}
for r in search_results
]
return await asyncio.to_thread(_format_sources)
async def _rerank_results(
self, question: str, search_results: List[Dict[str, Any]], skip_rerank: bool = False
) -> List[Dict[str, Any]]:
"""使用 Qwen3-Reranker 对搜索结果批量重排序"""
if skip_rerank:
self.logger.info("跳过重排序")
return search_results
if not search_results:
return []
# ==== Prompt 设置 ====
prefix = (
"<|im_start|>system\n"
"You are a helpful assistant that determines whether a document answers a given query. "
'Respond only with "yes" if the document is helpful, otherwise "no".\n'
"<|im_end|>\n"
"<|im_start|>user\n"
)
suffix = "<|im_end|>\n<|im_start|>assistant\n\n\n\n\n"
prefix_tokens = self.tokenizer.encode(prefix, add_special_tokens=False)
suffix_tokens = self.tokenizer.encode(suffix, add_special_tokens=False)
# 构造符合格式的输入
def format_pair(query: str, doc: str) -> str:
return f": {query}\n: {doc}"
pairs = [
format_pair(question, r["content"][:1000]) # 文本截断,避免超长
for r in search_results
]
# 分词 + 拼接前后缀 + padding
inputs = self.tokenizer(
pairs,
padding="max_length",
truncation="longest_first",
return_attention_mask=True, # 确保返回 attention_mask
max_length=8192 - len(prefix_tokens) - len(suffix_tokens),
return_tensors="pt",
)
# 手动添加前后缀
batch_size = inputs["input_ids"].shape[0]
max_len = 8192
# 创建新的输入张量
new_input_ids = torch.full((batch_size, max_len), self.tokenizer.pad_token_id, dtype=torch.long)
new_attention_mask = torch.zeros((batch_size, max_len), dtype=torch.long)
for i in range(batch_size):
# 获取原始序列(去除padding)
original_ids = inputs["input_ids"][i]
original_mask = inputs["attention_mask"][i]
actual_length = original_mask.sum().item()
# 构建新序列:prefix + original + suffix
new_sequence = (
prefix_tokens + original_ids[:actual_length].tolist() + suffix_tokens
)
new_length = len(new_sequence)
if new_length <= max_len:
new_input_ids[i, :new_length] = torch.tensor(new_sequence)
new_attention_mask[i, :new_length] = 1
inputs = {
"input_ids": new_input_ids.to(self.rerank_model.device),
"attention_mask": new_attention_mask.to(self.rerank_model.device),
}
# 获取 yes / no 的 token id(初始化时保存也可)
token_true_id = self.tokenizer.convert_tokens_to_ids("yes")
token_false_id = self.tokenizer.convert_tokens_to_ids("no")
# 推理评分
self.logger.info("模型准备输入完毕,开始推理...")
with torch.no_grad():
outputs = self.rerank_model(**inputs)
logits = outputs.logits
# 检查 logits 的维度
if logits.dim() == 3: # 如果是3维,取最后一个token的logits
logits = logits[:, -1, :]
elif logits.dim() != 2: # 如果是2维,直接使用
raise ValueError(f"Unexpected logits shape: {logits.shape}")
# 提取 yes/no token 的 logits
true_logits = logits[:, token_true_id]
false_logits = logits[:, token_false_id]
# 推荐用 logits 差值作为分数
scores = (true_logits - false_logits).tolist()
self.logger.info("模型推理完成")
# 写入每条结果
for r, score in zip(search_results, scores):
r["rerank_score"] = round(float(score), 4)
self.logger.info(f"重排序完成,得分范围: {min(scores)} - {max(scores)} \n\n {json.dumps(search_results, indent=4)}")
return search_results
async def _build_context_async(self, search_results: List[Dict[str, Any]]) -> str:
"""异步构建上下文"""
def _build_context():
return "\n\n".join(
[
f"文档片段 {i+1} (来源: {result['metadata']['filename']}):\n{result['content']}"
for i, result in enumerate(search_results)
]
)
return await asyncio.to_thread(_build_context)
async def _format_sources_async(
self, search_results: List[Dict[str, Any]]
) -> List[Dict[str, Any]]:
"""异步格式化源信息"""
def _format_sources():
return [
{
"filename": result["metadata"]["filename"],
"content": (
result["content"][:200] + "..."
if len(result["content"]) > 200
else result["content"]
),
"similarity": 1 - result["distance"],
}
for result in search_results
]
return await asyncio.to_thread(_format_sources)